Research Idea Convergence

Installation
SKILL.md

Research Idea Convergence

Canonical Summary

Interactive research direction selection for academic projects. Generates multiple candidate research directions from survey/literature results, presents them with structured comparison for user selection, then converges to a final publishable angle.

Trigger Rules

Use this skill when:

  • The project enters the ideation stage after completing the survey stage
  • The user asks about research direction, publishable angle, or idea selection
  • The current task involves generating or selecting a research direction

Do NOT use this skill when:

  • The user has already decided on a specific research direction and just wants to refine it
  • The project is still in the survey stage (use survey skills instead)

Execution Contract

  • Resolve every relative path from this skill directory first.
  • Do not write generated artifacts back into the skill directory; save them inside the active project workspace.
  • Never skip the user selection checkpoint — this is the core purpose of this skill.

Working Rules

  1. Read existing survey outputs before generating candidates.
  2. Generate 2-4 candidate directions, no more, no fewer.
  3. Each candidate must be concrete enough to evaluate (not vague platitudes).
  4. [USER_CHECKPOINT] — Always stop and wait for user selection. Never auto-select.
  5. Only after user confirmation, write the final direction to output files.

Step-by-step Instructions

Step 1 — Read Survey Context

Read all available survey outputs to understand the research landscape:

File Purpose
.pipeline/docs/research_brief.json Research topic, goal, target venue
.viewerleaf/research/Survey/reports/*.md Literature screening, gap summary
.pipeline/tasks/tasks.json Current task states and progress

Extract:

  • Research domain and specific problem area
  • Key gaps identified in the literature
  • Strong baselines that exist (traditional and learning-based)
  • Available datasets / experimental platforms
  • Target venue constraints (if specified)

Step 2 — Generate Candidate Directions

Produce 2-4 candidate research directions. For each candidate, provide:

## 候选方向 N:[方向名称]

**核心思路:** [1-2 句技术路线和创新点]

**研究问题:** [具体要回答的科学问题]

**方法概要:**
- [关键技术 1]
- [关键技术 2]
- ...

**对标基线:**
- [需要对比的已有方法 1]
- [需要对比的已有方法 2]

**目标刊物:** [适合投稿的会议/期刊]

**✅ 优势:**
- [优势 1]
- [优势 2]

**⚠️ 风险:**
- [风险 1]
- [风险 2]

**可行性评分:** ⭐⭐⭐⭐☆ (x/5)
- 实验资源: x/5
- 时间可控: x/5
- 技术难度: x/5
- 发表潜力: x/5

Candidate Generation Principles

  1. 差异化 — 候选方向之间应有明显区别(不同技术路线、不同创新角度、不同应用场景)
  2. 可比性 — 使用统一的评估维度,方便用户横向对比
  3. 务实性 — 考虑用户的实际资源(本科 vs 硕博、实验周期、计算资源)
  4. 梯度风险 — 包含保守稳妥的方向和激进创新的方向,让用户在风险与收益之间选择

Step 3 — Present Comparison Table

在列出所有候选方向后,给出一个汇总对比表:

## 📊 候选方向对比

| 维度 | 方向 1 | 方向 2 | 方向 3 |
|------|--------|--------|--------|
| 创新性 | ⭐⭐⭐ | ⭐⭐⭐⭐ | ⭐⭐⭐⭐⭐ |
| 可行性 | ⭐⭐⭐⭐⭐ | ⭐⭐⭐ | ⭐⭐ |
| 发表潜力 | ⭐⭐⭐ | ⭐⭐⭐⭐ | ⭐⭐⭐⭐⭐ |
| 实验周期 | 2 周 | 4 周 | 6 周 |
| 推荐度 | 🟡 稳妥 | 🟢 推荐 | 🔴 激进 |

Step 4 — USER CHECKPOINT ⛔

[USER_CHECKPOINT] — 必须在此处停下,等待用户回复。

向用户提问:

  1. 你倾向于哪个方向?
  2. 是否需要组合某些方向的元素?
  3. 有没有你自己的想法想要加入?
  4. 对某个方向的风险有没有顾虑?

绝对不要跳过这一步。不要替用户做选择。

等用户明确选择后,再进入 Step 5。

Step 5 — Converge and Write

用户确认后:

  1. 写入 publishable_angle.md

    路径: .viewerleaf/research/Ideation/publishable_angle.md
    

    包含:选定方向的完整描述、用户的修改意见(如有)、对标基线、预期贡献

  2. 更新 research_brief.json

    路径: .pipeline/docs/research_brief.json
    

    更新 briefGoalstageNotes.ideation

  3. 更新 tasks.json

    路径: .pipeline/tasks/tasks.json
    

    标记 ideation 任务完成,触发下一阶段任务

  4. 记录选择理由publishable_angle.md 末尾附加:

    ## 选择记录
    - 候选方向数:N 个
    - 用户选择:方向 X [+ 用户修改内容]
    - 选择理由:[用户给出的理由或综合分析]
    - 弃选方向:方向 Y (原因)、方向 Z (原因)
    

Output Templates

publishable_angle.md

Use templates/publishable_angle.template.md:

# Publishable Angle

## 研究方向
[选定方向名称]

## 核心创新点
[1-2 句话概括]

## 研究问题
[具体要回答的科学问题]

## 方法概要
- ...

## 对标基线
- ...

## 预期贡献
- ...

## 目标刊物
[期刊/会议名称]

## 选择记录
- 候选方向数:N 个
- 用户选择:方向 X
- 选择理由:...

Integration with Other Skills

Upstream This Skill Downstream
research-literature-trace (survey reports) → 读取 gap/baseline 信息
research-pipeline-planner (brief) → 读取研究主题和目标
生成候选 → 用户选择 → 写入 inno-idea-eval (评审选中方向)
inno-experiment-dev (实验设计)

Notes for LLMs

  1. 核心原则:这个 skill 的价值在于交互,不在于生成。生成候选只是手段,让用户做出知情选择才是目的。
  2. 不要跳过 USER_CHECKPOINT:即使用户看起来已经有偏好,也要完整展示所有候选并明确收集确认。
  3. 不要过度推荐:可以标注"推荐"但不要暗示只有一个正确答案。每个方向都有其适用场景。
  4. 尊重用户原有思路:如果用户已有初步想法,把它作为候选之一纳入对比,而不是忽略。
  5. 差异化很重要:不要生成几个只有微小差别的方向,那等于没给选择。技术路线、创新角度、目标场景应有明显区分。
Related skills
Installs
GitHub Stars
455
First Seen